2026-10-03 · Agentic Realism
ChatGPT and patient data: what you can paste, what you cannot
Three zones for using AI in a practice without breaching professional confidentiality
Session notes, an insurance report, a patient’s message: none of it goes into a consumer assistant. What is about no one, yes. Three zones, an anonymisation checklist, the settings to check — and the point where you need a supervised tool rather than a personal account.
The short answer
In a consumer assistant such as ChatGPT or Claude, paste nothing that would let a patient be recognised: no session notes, no report, no message received. What contains no personal data can go in without concern; a properly anonymised situation can go in with care. For regular use on real files, you need a tool framed by a contract, not a personal account.
The situation
It is 8 p.m., the insurance report must go out tomorrow, and your notes fill two handwritten pages. The assistant open in the other tab writes quickly and well. You would only need to paste the notes and ask for a clean summary.
That is exactly the gesture to avoid. Not because the tool is malicious, but because you have just passed health data that identifies a person to a company with which neither you nor your patient has a suitable contract — and often outside Switzerland. The good news: much of what takes up your time can be handed to AI without ever touching a file.
Why it is sensitive
Professional confidentiality. For many health professions — doctors, psychologists, physiotherapists, osteopaths, among others — revealing information entrusted in the course of practice is a criminal offence in Switzerland. For other therapists, discretion remains an obligation towards the patient and, often, towards your professional association.
Data protection. Health data is sensitive data under the Swiss Federal Act on Data Protection. Handing it to a third party, let alone abroad, requires a solid basis — a contract, guarantees, informing the patient — that a free sign-up to an online service does not provide.
This is not legal advice: the exact rules depend on your profession and your canton. But the practical line is simple.
Three zones
- GREENNo personal dataYour website texts, practice FAQ, a blank letter template, rewording a public document, general research.
- ORANGECarefully anonymisedA situation rewritten for supervision or a question of method, with nothing that would allow the person to be recognised.
- REDA recognisable personSession notes, a named report, an insurance report, a patient’s message, a diary with reasons for consultation.
Green: everything that is about no one
This is where most of the gain lies, and it is often underestimated: writing the “My approach” page of your website, preparing an FAQ for new patients, creating a blank referral-letter template, rewording a popular-science text, preparing a presentation, sorting your ideas for a training course. No patient data, no risk on that side.
Orange: a situation rewritten so no one is recognised
Preparing a supervision, thinking through a question of method, looking for reading leads from a situation: it is possible, provided you truly anonymise. Removing the name is not enough — it is the combination of details that identifies.
| Element | What to do |
|---|---|
| Name, first name, initials | Remove — initials often identify |
| Precise dates | Replace with durations: “for three months” |
| Exact age | Replace with a range: “in their forties” |
| Place, employer, school | Remove or make general |
| Rare occupation, unusual family situation | Generalise |
| Rare diagnosis, striking event | Ask whether the case remains recognisable |
| Combination of what remains | Re-read: could someone close recognise the person? |
The final test is simple: if someone close to the patient read the text, could they recognise them? If the answer is “maybe”, the text stays red.
Red: everything about a recognisable person
Session notes, reports, insurance reports, letters to a doctor, messages received from a patient, a diary with reasons for consultation, and even a simple list of names with appointment dates. None of it goes into a personal account of a consumer assistant, whatever its settings.
Settings to check, even in the green zone
Model training. Depending on the service and the plan, your conversations may be used to improve the models. Consumer versions have a setting to refuse this; check that it is switched off. Business offers generally do not use them by default.
One account for the practice. Avoid using the same account for your private life and your practice: the history gets mixed, and it becomes hard to know what was put there.
History. Deleting a conversation does not always erase it immediately from the provider’s servers. The only reliable protection is not to put in what should not be there.
When a consumer tool is no longer enough
If you want an assistant to genuinely prepare the practice’s admin — appointments, reminders, intake documents, invoices — it has to see names and dates. The question then changes in nature: it is no longer “what can I paste?”, but “where is the data processed, under what contract, with what access?”.
A supervised system answers these questions before starting: hosting in Switzerland or Europe, documented providers, access limited to what is strictly needed, a log of what the assistant did, and a clear boundary around clinical content, which stays out of its reach. That is the approach described in Privacy is not a slogan, and the one we apply for therapists and clinics.
What it costs
Applying the three zones costs nothing. If you want to learn to use these tools on your own tasks — with confidentiality rules written for your practice — a solo training session costs CHF 350 for two hours, with a personalised memo. A supervised assistant for the practice’s admin falls under the agentic office, from CHF 490 for a first brick. You only pay if you approve the proposal.
The limits
This article gives a cautious line of conduct, not a legal analysis of your situation. If you are considering regular use of AI on patient data, talk to your professional association or a specialised lawyer before choosing a tool.
And even in the green zone, the assistant makes mistakes: it invents references, oversimplifies, takes a tone that is not yours. Everything it produces is re-read before use.
Frequently asked questions
Is the patient’s consent enough?
It does not settle everything. Consent must be free, informed and specific: the patient should know which service, in which country, under what conditions. You remain responsible for that choice. For a consumer tool, the cautious course is not to rely on consent.
Does a business plan of ChatGPT or Claude settle the question?
It improves the situation — no training on your data, a contract, guarantees — but does not on its own settle the questions of professional confidentiality and transfer abroad. Read the contract, and seek advice if you put real files into it.
Can I use AI to write to a patient?
Yes to prepare a message template — appointment confirmation, practical information, reminder — written without personal data, which you then complete yourself. No to having it read and summarise what the patient wrote to you.
Going further
If your patients first find you on Google, the Google profile checklist also explains how to reply to reviews without breaching professional confidentiality. And the 30-minute diagnostic is free, if you want to know what in your practice can be delegated without risk.